Every month, someone in my community gets a government notice they can't fully understand — not because they can't read, but because official language is its own dialect. A property tax notice, a ration card renewal, a health advisory — all written in dense, formal Bengali or Telugu that buries the one thing that actually matters: what you need to do, and by when.
That's the problem my team set out to solve at the WeMakeDevs Bharat Builds Tour "First Commit" hackathon: Notice, Made Simple — an app that takes a photo or pasted text of a Bengali or Telugu notice and returns a plain-language summary, clear action steps, the deadline, and an audio version you can just listen to.
What I built
As team lead and frontend developer, I owned the entire React 19 + Vite + Tailwind CSS frontend — the language picker, the upload/paste-text flow, the processing states, and the result screen with its simplified summary, category badges, deadline chips, and audio player. I deployed it on AWS Amplify, which auto-builds and deploys straight from GitHub on every push — genuinely the smoothest part of the whole weekend once I got the build output directory set correctly for Vite.
Where AWS came in
Our backend runs on AWS Lambda behind Amazon API Gateway, with Amazon S3 storing uploaded images and generated audio, and Amazon DynamoDB persisting results. For the AI simplification itself, we integrated Amazon Bedrock with Claude Sonnet 4.5 — and once model access cleared, the output quality for Bengali/Telugu text was genuinely impressive: clean, structured JSON with no hallucinated details.
What actually happened (the messy part)
Here's the part most hackathon writeups skip: we hit real account-level walls. Amazon Textract, it turns out, doesn't support Bengali or Telugu script at all — a discovery that reshaped our whole OCR strategy mid-build. We got INVALID_PAYMENT_INSTRUMENT and SubscriptionRequiredException errors that looked like permission problems but were actually billing verification gates — I nearly recommended granting broad AdministratorAccess to fix it before realizing the real issue was a missing payment method on the account.
Rather than stall, we built resilience in: Tesseract OCR bundled directly into the Lambda for Bengali/Telugu image text, Microsoft Edge TTS for audio (since Polly doesn't cover these languages), and a Groq-based LLM fallback so the app kept working end-to-end while our AWS account access finished clearing.
What I learned
The biggest lesson wasn't technical — it was that "working" beats "perfect" under a deadline. We shipped a real, deployed, end-to-end product that solves a genuine problem, even while parts of it ran on fallback infrastructure instead of the AWS services we originally designed for. Documenting those limitations honestly, rather than hiding them, felt like the right call — and it's something I'd do again.
Loading article